[stochastic_monk]

Transformer-XL uses recurrence, and most NLP SOTA is still with LSTMs. I’m not sure I’d expect attention mechanisms to fully replace recurrence.